morluto/rea leads the week: give your coding agent a website, an Electron app, a .NET assembly or a native binary, and it comes back with a readable reconstruction in which every claim carries the evidence it came from and the limits behind it. storytold/photocraft follows as the clean-room Photoshop rebuilt in pure Rust, and boykopovar/AnyPS5 takes third by relinking PS5 executables into native Linux or Windows programs, no emulation layer involved.
The rest of the storytold suite fills much of the chart. storytold/artcraft blocks out 3D scenes before generating images or video, storytold/lightcraft reimplements Lightroom with non-destructive raw edits, storytold/filmcraft rebuilds the Premiere Pro timeline, color panel and audio stack, and storytold/pdfcraft appends every edit while the original PDF bytes stay untouched. storytold/wordcraft brings Microsoft Word with track changes, storytold/vectorcraft redraws Illustrator with live blends and a perspective grid, and storytold/effectcraft closes the set with After Effects compositions, built-in Lottie export and a command registry that the CLI and MCP server share.
Agent working habits get their own cluster. mattpocock/skills gives your coding assistant the discipline of asking detailed questions before it builds the wrong thing, while DietrichGebert/ponytail pushes the opposite direction, forcing agents to reuse existing code and justify every new abstraction. Ebony-Vinyl/dsh-our-free-model fills the DeepSeek Harness model picker with free frontier models pulled from public gateways, no login or key needed. cathrynlavery/diagram-design turns explanations into HTML and SVG diagrams that inherit a publication’s colors and fonts, and Panniantong/Agent-Reach wires command-line agents into reading Twitter, Reddit, YouTube and more, with channel diagnostics that show what is connected and what needs configuration.
Research and self-hosted models meet in the middle of the chart. openai/math is the week’s curiosity: 722 math manuscripts and proof artifacts written by an internal OpenAI model while it was tested on open research problems, from the irrationality exponent of pi to the 3D Vlasov-Maxwell system, many carrying Lean formalizations you can check yourself. Niko1221/Strata answers the hardware side, installing the 125-billion-parameter Qwen3.8-Flash-Next on a single 12 to 24 GB gaming GPU and serving an OpenAI-compatible API locally.
Verification and repetition are the theme of the next group. tester-army/e2e lets you describe a test goal in plain language while an agent drives your web or mobile app to reach it, then records verified steps for replay without model calls. stablyai/orca runs several terminal coding agents in separate Git worktrees so attempts at the same task can be compared side by side, and lexmount/moli is a Rust headless browser that hands agents page structure instead of pixels. robbietilton/Compositor is a free macOS image editor built around the Photoshop compositing workflow, with layers, adjustment layers and content-aware fill.
The remaining entries leave the terminal. eternity4719/HowToLiveBetter holds fourth with its evidence-graded Chinese handbook of everyday advice, from health and rent disputes to scam avoidance, each entry priced by cost, return and strength of evidence. DuarteSantos8/openGym is a self-hosted workout tracker that imports history from FitNotes, Strong and Hevy, debpalash/VoiceStudio keeps voice cloning and transcription local, and shihabal3amri/DiPlay puts CarPlay on compatible BYD head units, with the author explicit that it is not Apple-certified. If you only try one this week, pick by your bottleneck: unverifiable agent claims point to rea, diagram chores point to diagram-design, and anyone in a BYD should read DiPlay’s compatibility notes before installing anything.